Hematology Blood Analysis with Computer Vision

The brief

Manual differential counts on blood films are slow and vary between operators. The goal was a vision pipeline that pre-classifies cells so laboratory staff confirm results rather than counting from scratch — decision support, not diagnosis.

What I did

• Built an ingestion step for microscope images, normalising illumination, colour and scale so slides captured in different sessions are comparable.
• Segmented cells from the background and separated touching cells before classification, since overlaps are where naive pipelines fail.
• Trained a convolutional classifier over labelled cell types, augmenting for stain and lighting variation so it generalises beyond one scanner.
• Set a confidence threshold: anything the model is unsure of is flagged for human review instead of being silently classified.
• Built a review screen where staff confirm or correct each call, and captured every correction as future training data.
• Kept the pipeline reproducible — versioned datasets, versioned models, and a held-out set to measure drift over time.

Outcome

• Analysts confirm and correct rather than counting from zero
• Low-confidence cells are escalated rather than quietly guessed
• Corrections feed straight back into the next training round

A qualified professional reviews and signs off every result; the model assists that judgement, it does not replace it.